A Complete Normal Model Problem-Solving Workshop

From a Contextual Question to a Normal Probability

A normal model problem often begins in ordinary language: What fraction of containers fall below a target? How likely is a measurement to lie between two limits? What value marks the 90th percentile? Solving these questions reliably takes more than a table lookup. First identify the measured quantity and the model; then translate the question into an event, calculate its probability or cutoff, and return to the original units to interpret the result.

The model describes a variable in a specified population or process. A normal calculation is only as appropriate as that model: a numerical answer does not establish that real measurements follow a normal distribution. The model should be supported by the context and, when data are available, by suitable checks such as a normal probability plot.

Definition: A normal model for a quantitative random variable $X$ specifies that $X\sim N(\mu,\sigma^2)$, where $\mu$ is the model mean and $\sigma$ is the model standard deviation. The distribution is symmetric about $\mu$, and its values are measured in the same units as $X$. The standard deviation is positive. The standardized variable $Z=(X-\mu)/\sigma$ has the standard normal distribution, with mean 0 and standard deviation 1.

For calculations, let $\Phi(z)=P(Z\leq z)$ denote the cumulative probability to the left of $z$ for a standard normal variable $Z$. A cumulative standard normal table gives values of $\Phi(z)$; software can calculate the same areas. Since a normal variable is continuous, the probability of one exact value is zero, so $P(X\leq c)=P(X<c)$ and $P(X\geq c)=P(X>c)$.

A Reliable Problem-Solving Sequence

Before using a table or software, write down the model and state the event of interest. A cutoff in original units becomes a standardized cutoff through $z=(c-\mu)/\sigma$. This conversion tells how many standard deviations the cutoff lies above or below the mean. For an interval, convert both endpoints, preserving their order.

Define the variable and model.
Say what one value of $X$ represents, give its units, and state $X\sim N(\mu,\sigma^2)$. Check that a normal model is reasonable for the context and that its assumptions are not contradicted by known information.

Translate the question into an event.
For example, “at least 488 grams” becomes $X\geq488$; “between 36 and 51 minutes” becomes $36\leq X\leq51$. The direction and both endpoints matter.

Standardize the cutoff or cutoffs.
For each original-scale cutoff $c$, calculate $z=(c-\mu)/\sigma$. Keep enough digits in intermediate results to avoid avoidable rounding error.

Calculate and check the probability or quantile.
Use a cumulative table, software, or a percentile lookup as appropriate. Check that probabilities lie between 0 and 1 and that the result is consistent with the event and the location of its cutoffs.

Interpret in context.
State the probability or proportion the model assigns to the event, name the measured quantity, and include its units when reporting a cutoff or measurement range.

Why Standardizing Works

Standardizing does not change which outcomes are being counted; it expresses them on a common scale. Subtracting $\mu$ shifts the center of the model to 0, and dividing by $\sigma$ measures distance in standard deviations. Because $\sigma>0$, division preserves the ordering of values. Thus an event involving $X$ can be rewritten as an event involving $Z$, after which a standard normal table or calculation can be used.

Justification: If $X\sim N(\mu,\sigma^2)$ and $Z=(X-\mu)/\sigma$, then the normal distribution of $X$ is shifted by $\mu$ and scaled by $\sigma$, giving $Z\sim N(0,1)$. For any cutoff $c$, $$ P(X\leq c) =P\left(\frac{X-\mu}{\sigma}\leq\frac{c-\mu}{\sigma}\right) =P\left(Z\leq\frac{c-\mu}{\sigma}\right) =\Phi\left(\frac{c-\mu}{\sigma}\right). $$ The inequality direction stays the same because $\sigma$ is positive. For an interval $a\leq X\leq b$, subtract the probability to the left of $a$ from the probability to the left of $b$.

Worked Examples

Worked Example: Checking a Lower Fill Limit

Hypothetical situation. A filling process is modeled as producing container contents with $X\sim N(500,8^2)$, where $X$ is the content in grams. What proportion of containers does the model predict will contain less than 488 grams?

Model check. The model is stipulated for this hypothetical process. Its mean is 500 grams and its standard deviation is 8 grams, so the lower limit is interpretable relative to the modeled variation. Container contents are positive, while this normal model technically allows negative values; zero is 62.5 standard deviations below the mean, so the model assigns a negligible probability to negative contents. In an actual application, the model’s fit would need to be assessed from process knowledge and data.

The event is $X<488$. Standardize the cutoff:

z=\frac{488-500}{8}=-1.5.

A cumulative standard normal table gives $\Phi(-1.50)\approx0.0668$. Equivalently, software can calculate the left-tail probability directly, for example with pnorm(488, mean = 500, sd = 8) in R.

P(X<488)=P(Z<-1.5)=\Phi(-1.5)\approx0.0668.

The model predicts that about 0.0668, or 6.68%, of containers have contents below 488 grams. This is a probability under the stated model, not a guarantee that exactly this percentage will fall below the limit in any particular finite batch.

Worked Example: Probability Between Two Time Limits

Hypothetical situation. A machine cycle time $X$ is modeled by $X\sim N(42,6^2)$, with times measured in minutes. Find the probability that a cycle takes between 36 and 51 minutes, inclusive.

Model check. Assume the cycle-time distribution is adequately represented by the stated normal model. The mean is 42 minutes and the standard deviation is 6 minutes. Although time cannot be negative, zero is 7 standard deviations below the mean, so negative times have negligible probability under this model. If this were a real process, the normal shape and the relevance of these parameters would need support from process knowledge or data.

The event is $36\leq X\leq51$. Convert each endpoint to a $z$-score:

z_{36}=\frac{36-42}{6}=-1, \qquad z_{51}=\frac{51-42}{6}=1.5.

The probability between the cutoffs is the cumulative area to the left of 1.5 minus the area to the left of $-1$:

\begin{aligned} P(36\leq X\leq51) &=P(-1\leq Z\leq1.5)\\ &=\Phi(1.5)-\Phi(-1)\\ &\approx0.9332-0.1587\\ &=0.7745. \end{aligned}

The model assigns probability about 0.7745, or 77.45%, to a cycle time between 36 and 51 minutes. Because the distribution is continuous, including or excluding either endpoint does not change the probability. As a quick check, the upper endpoint is 1.5 standard deviations above the mean and the lower endpoint is one standard deviation below it; a probability greater than one-half is therefore plausible.

Software can compute the same area as the difference of two cumulative probabilities, for example pnorm(51, mean = 42, sd = 6) - pnorm(36, mean = 42, sd = 6) in R. Entering the larger cutoff first ensures that the difference is nonnegative.

Worked Example: Finding a 90th-Percentile Lifetime

Hypothetical situation. A battery lifetime $X$ is modeled as $X\sim N(1200,90^2)$, where lifetime is measured in hours. Find the model’s 90th-percentile lifetime.

Model check. Assume the normal model is suitable for this hypothetical battery population. Its standard deviation is 90 hours. The model technically permits negative lifetimes, but zero is about 13.33 standard deviations below the mean, so the probability it assigns to negative values is negligible. A real-world use would require checking the model against relevant data; specifying a normal model alone does not verify it.

The 90th percentile is the value $x_{0.90}$ for which $P(X\leq x_{0.90})=0.90$. The corresponding standard normal quantile is approximately $z_{0.90}=1.2816$. Convert this standardized value back to hours:

\begin{aligned} x_{0.90} &=\mu+z_{0.90}\sigma\\ &=1200+(1.2816)(90)\\ &=1200+115.344\\ &\approx1315.3\text{ hours}. \end{aligned}

Under the model, approximately 90% of battery lifetimes are at or below 1315.3 hours, and approximately 10% are above it. This is a model percentile, not a claim that each battery has a 90% chance of lasting exactly that long. Software can find the cutoff directly, for example with qnorm(0.90, mean = 1200, sd = 90) in R.

Tables, Software, and Sanity Checks

A cumulative normal table and software both provide areas to the left of a standardized value or find a standardized value for a given cumulative area. When using a table, first identify whether its entries are cumulative left-tail areas; do not assume every table is arranged the same way. If the table reports $\Phi(z)$, a right-tail probability is $1-\Phi(z)$, and a probability between two cutoffs is the difference of their cumulative areas. Software can evaluate these expressions directly, but the event and model still need to be specified correctly.

Before accepting a numerical answer, use checks tied to the question:

A probability must be between 0 and 1. A negative interval probability usually signals that the cumulative areas were subtracted in the wrong order.
If the cutoff is below the mean, the probability of being below it must be less than 0.5; if it is above the mean, that probability must be greater than 0.5.
For a right-tail event, use the area above the cutoff, not the area below it. A cutoff far above the mean should not produce a large upper-tail probability.
A percentile cutoff above the mean must correspond to a percentile above 50. When converting back to original units, multiply the $z$-value by $\sigma$, then add $\mu$.
Report a probability as dimensionless (or as a percentage), but give a cutoff or measurement in the original units.

These checks do not establish that a normal model is valid. They identify calculation or interpretation errors conditional on the model being appropriate.

Common Mistake: Reversing the Tail or Losing the Context

A frequent error is to look up the cumulative area to the left of a cutoff and report it when the question asks for values above that cutoff. For example, if a cutoff has $z=1.5$, $\Phi(1.5)\approx0.9332$ is the probability below it. The probability above it is $1-0.9332=0.0668$. The correction is to write the event first and check whether it describes a left tail, right tail, or interval before selecting the table entry.

Another mistake is to report a standardized cutoff as though it were the measurement. A $z$-score of 1.2816 is a distance in standard deviations, not a lifetime of 1.2816 hours. To return to the measurement scale, use $x=\mu+z\sigma$ and retain the original units. Finally, a calculated probability should not be interpreted as certainty about the next observation: it describes the proportion of the modeled distribution in the specified event.

Key takeaway: Define $X$ and its normal model, translate the wording into an event or percentile, standardize or unstandardize as needed, check the result, and state what it means in context. The arithmetic is useful only when both the model and the interpretation match the question.

Check Your Understanding

For each question, identify the event or target before calculating.

If $X\sim N(70,5^2)$, write the $z$-score for the cutoff $x=63$. Is the probability below 63 greater or less than 0.5?
A normal model has cumulative area 0.92 to the left of a cutoff. What is the probability above that cutoff?
For $X\sim N(100,12^2)$, express $P(88\leq X\leq112)$ as a difference of standard normal cumulative probabilities.
A value is 1.4 standard deviations above a mean of 250, with standard deviation 20. Convert it to the original measurement scale and include units of your choice.
Why does a technically correct normal calculation not, by itself, show that the model is appropriate for real measurements?